Progress in the research of artificial intelligence in andrology: a narrative review

Introduction

Background

Andrology is an emerging medical discipline that mainly studies the structure, function, and pathological changes of the male reproductive system. However, in clinical practice, the diagnosis and treatment of male diseases are hindered by some long-standing problems. For example, traditional semen laboratory analysis heavily relies on the operator’s experience, which brings considerable subjectivity and inter-observer variability. The diagnosis of sexual dysfunction mainly relies on patients’ self-reports or questionnaire evaluations. This process lacks objective biomarkers or indicators to support clinical decisions. The pathogenesis of many prostatic diseases remains unclear. This situation further increases the difficulty of precise diagnosis and personalized treatment. On the other hand, the data generated by andrological surgery and emerging technologies are complex and multimodal, which are difficult to integrate and interpret by traditional methods. The clinical situation described above indicates that new technical solutions are needed.

Artificial intelligence (AI) is a branch of computer science. It aims to develop theories, methods and applications that can simulate, extend and enhance human intelligence. In recent years, AI has developed rapidly, especially with advancements in subfields such as machine learning (ML). These advancements have provided new possibilities for solving problems such as medical diagnosis, risk prediction and treatment optimization (1).

Rationale and knowledge gap

Existing literature reviews have pointed out that through algorithms and physical medical devices, AI can generate social benefits. At the same time, establishing scientific ethical standards is also necessary (2). Currently, most published reviews only focus on a single disease entity or are limited to the technical aspects of the AI methodology. This situation makes it difficult for clinicians and researchers to understand the role of AI in the entire field of male sexual disorders. Additionally, the limitations of AI applications in the field of andrology have not been systematically and fully explored. Important issues such as data quality defects, insufficient model interpretability, and ethical problems are often only briefly mentioned and rarely subjected to critical analysis.

Objective

To address these deficiencies, this review systematically summarizes and critically assesses the progress of AI in the application to male infertility (MI), erectile dysfunction (ED), premature ejaculation (PE), and prostatic diseases. At the same time, this article systematically summarizes specific development strategies. It aims to make doctors and patients fully understand and use AI better. We present this article in accordance with the Narrative Review reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0474/rc).

Methods

We searched the literature for this narrative review, conducting a comprehensive search in four authoritative electronic databases: PubMed, Web of Science, Scopus and China National Knowledge Infrastructure (CNKI). The search period was from 2010 to 2026 to focus on classic literature and the latest progress in this field. Search terms included AI, ML, MI, ED, PE, and prostatic diseases.

Firstly, the basic definitions and epidemiological data were obtained by searching the above terms separately. Then, medical keywords were integrated and Boolean operators (AND, NOT, OR) were used to retrieve the results precisely. The search process was conducted independently by two researchers, first reviewing titles and abstracts, followed by full-text assessment. Disagreements were discussed and resolved by a third researcher. The search language included Chinese and English, and the types of included articles included reviews, original studies, and case reports. Duplicate published, non-peer-reviewed, and off-topic articles were excluded.

The purpose of this search was to comprehensively grasp the research progress of AI in the field of andrology. To comprehensively pay attention to the latest situation and limit by space, we focused on the recently published literature and mainly researched in imaging, large language models (LLMs), and risk prediction, considering the characteristics of the disease (Table 1).

The core technologies, common challenges, and corresponding solutions of AI

Main learning paradigms and their limitations

There are some limitations in andrology. To address these limitations, AI intervention offers a new technical approach to overcome the shortcomings in traditional diagnosis and treatment. The deep application of AI in andrology essentially involves simulating and extending clinical practice through its subfield, ML. These core paradigms can be categorized as supervised learning, unsupervised learning and reinforcement learning (Figure 1A-1C) (3).

Figure 1 Major machine learning paradigms and clinical challenges and solutions. (A) Supervised learning learns patterns by training on data that comes with clear labels or features. (B) Unsupervised learning discovers the hidden structures and patterns within the data, and clusters the unlabeled and unordered data into groups with similar characteristics. (C) Reinforcement learning is a process where an intelligent agent performs tasks in a specific environment and learns through a continuous feedback mechanism to acquire the optimal behavior strategy. (D) Single-center database and population selection bias. (E) Algorithm black box, with an opaque decision-making process. (F) Large sample multi-center data can enhance the representativeness of the data and the generalization ability of the model. (G) Improve the interpretability of the model. (H) There is a two-way positive interaction between medical practitioners and artificial intelligence. (The image was created at BioGDP.com).

Supervised learning relies on training with labeled datasets to map complex clinical features to predictable outputs. This approach can be used for disease prediction, image diagnosis, and drug development. Unsupervised learning is the process of exploring the inherent structure and distribution of data when there are no prior labels. In the field of andrology, it can be used to identify new disease subgroups or potential biomarkers from multi-dimensional clinical or omics data, providing clues for understanding the complex causes.

Reinforcement learning is a learning process where an agent interacts with an unknown dynamic environment. This method has potential in developing personalized treatment plans and optimizing surgical procedures. It is particularly applicable to complex clinical scenarios that require multi-step decision-making (4).

Andrological diseases have their own particularities. The patient is under both physical and psychological pressure and is too embarrassed to speak out. When seeking medical treatment, they tend to conceal their actual condition. Under these particularities, obtaining high-quality clinical data becomes difficult. Therefore, AI faces two main challenges in the practice of andrology, namely the challenges at the data level and the challenges at the model level (Figure 1D,1E).

Technological risks and countermeasures

At the data level, the absence of large-scale and multi-institutional databases will pose obstacles to the development of AI models. This deficiency will also limit the generalization ability of the models in different clinical settings. On the other hand, if the training data does not fairly represent different populations, it will generate biases. Such biases pose a challenge for developing fair and non-discriminatory models (5). At the model level, some high-performance deep learning models have complex decision-making logic. This logic makes it difficult to explain the models. As a result, the acceptability and practicality of these models in medical diagnosis and detection have also declined (6).

To address these challenges, data augmentation is a feasible technical strategy. In the field of medical imaging, Garcea et al. demonstrated through research that data augmentation can enhance the performance of models in tasks such as classification and segmentation (7). Traditional data augmentation methods often destroy the biological features of images. To address this issue, Wu et al. proposed a new solution: by dimension expansion, the data utility features are fused with the original image, and at the same time, a channel weight feature extractor is constructed. This method can expand the data and improve the performance of the model. At the same time, the interpretability of the model’s decisions has also been enhanced (8). In non-image domains, one study proposed a data augmentation method that is based on entity perception and employs a local fusion strategy. When the labeled medical text data is limited, model performance can be improved through this method (9). These studies collectively indicate that future work should make efficient use of data and ensure the transparency of model decisions. Through these approaches, AI will no longer merely be available but will instead become trustworthy in clinical work (Figure 1F-1H).

Applications of AI in andrological diseases

MI: the paradigm shift triggered by AI

Infertility is a common health issue in marriages. Approximately 15% of couples worldwide are affected by it. Among these cases, about half are related to male factors (10). AI has brought revolutionary changes to the precision diagnosis and treatment of MI by participating in semen analysis, assisting in reproductive technology, microsurgery and revealing the pathological mechanism.

Standardized semen parameter testing

Traditional semen analysis is regarded as the cornerstone of diagnosis. However, this method has several drawbacks: it is highly subjective, has poor repeatability, uses inconsistent evaluation criteria, and is time-consuming (11). The core value of AI in the assessment of semen parameters lies in achieving standardized, automated analysis. Systems based on deep learning models, such as convolutional neural networks, can provide a highly precise and objective interpretation of sperm concentration, motility and morphology (12). An algorithm developed by Yang et al. enables the real-time tracking of live, unstained sperm and can independently complete morphological analyses of the head, midpiece and principal piece. This facilitates the rapid, non-invasive screening of high-quality live sperm with good progressive motility and normal morphology, thereby improving assisted reproductive outcomes (11). Another study addressed the limitations of trajectory smoothing algorithms used in traditional computer-aided sperm analysis systems by enhancing the precision of sperm kinematic parameter calculations. They introduced discrete cosine transform smoothing and path average width metrics. Based on this, by combining the InceptionTime deep learning model, this goal was achieved (13). These advancements indicate that AI is steering semen analysis towards the direction of repeatability and quantification.

AI helps sperm integrity assessment

The sperm DNA fragmentation index (DFI) is a crucial indicator for evaluating male fertility. In clinical testing, sperm chromatin structure analysis (SCSA) based on flow cytometry is regarded as the “gold standard”. However, impurities in semen can interfere with the detection results of flow cytometry. For samples with azoospermia or oligospermia, this interference may produce false positive signals, thereby causing the DFI value to be artificially elevated. One study compared the AI fluorescence method with traditional flow cytometry. The research results showed that the AI fluorescence method had higher specificity in distinguishing sperm from semen impurities. For patients with moderate to severe oligospermia, the test results obtained by this method might be closer to their actual physiological state (14). Another study developed a multimodal optical system. This system can perform multiple images of sperm cells. These images are then analyzed by the MobileNet convolutional neural network to predict the DFI value of individual sperm cells. Through this method, the success rate of assisted reproduction has been improved (15).

Apart from SCSA, sperm chromatin dispersion (SCD) testing is also a widely used technique in clinical settings for assessing sperm DNA degradation, as it is cost-effective. However, the traditional SCD test mainly relies on manual microscopic assessment of the size of the halos, which is subjective, time-consuming, and prone to significant inter-observer differences. Hsu et al. developed an automated SCD diagnostic workflow by combining the LensHooke^® R10 system with the X12 AI analyzer, which significantly reduced the interpretation differences (16). Additionally, Kuroda et al.’s comparative study demonstrated that this AI-assisted halo assessment platform has a strong correlation with the “gold standard” SCSA. By eliminating the subjectivity of manual counting, especially in samples with limited sperm counts, this automated method significantly improved the reproducibility and efficiency of diagnosis (17).

In summary, AI has improved the efficiency of routine semen analysis and solved technical challenges in specific clinical situations, shifting the assessment of sperm DNA integrity from empirical to algorithm-driven standardization.

Risk prediction and omics study of MI

The application of AI has further expanded into the field of disease risk prediction, as well as the clarification of pathophysiological mechanisms. In optimizing clinical diagnostic workflows, Kobayashi et al. constructed an AI predictive model based on serum hormone levels. This model enables a preliminary assessment of MI risk and non-obstructive azoospermia, eliminating the need for semen analysis. Thus, it provides a new tool for rapid clinical screening (18). In exploring etiological mechanisms, a team used ML to integrate multi-omics data and discovered that obesity and MI share five core genes. They constructed an intervention network of compounds and traditional Chinese medicines, suggesting new approaches to the treatment of obesity-associated MI using traditional Chinese medicine (19).

Robot-assisted minimally invasive surgery

Moreover, a notable phenomenon is that AI robots are gradually being introduced into some microsurgical operations. Examples include varicocele resection and microdissection testicular sperm extraction (micro-TESE) (20). Physiological hand tremors can be effectively eliminated through these techniques. At the same time, the accuracy of suturing has been improved and the time required for training has been significantly shortened (21). In micro-TESE, Mohamed et al. developed an augmented reality microscope system combined with deep learning, which increased the sperm detection efficiency by approximately twice in simulated samples. However, it has not yet been clinically validated (22). Kandil et al. confirmed that the AI image classifier could accelerate the identification of rare sperm, but this was only a proof-of-concept and lacked prospective multicenter studies (23). Gül et al. used preoperative parameters to establish a random forest (RF) model for 470 men with Klinefelter syndrome, and it is one of the earliest tools in this population to have undergone external validation (24).

AI and assisted reproductive technologies

In the assisted reproductive field, Gao et al. conducted biochemical classification of sperm and found that specific subtypes were associated with a higher rate of blastocysts. However, this study had the limitation of using fixed sperm (25). Li et al. developed a model for predicting clinical pregnancy and found that body mass index (BMI), anti-Müllerian hormone, follicle-stimulating hormone, and female age were the main predictors (26). Both studies were in the exploratory stage and had not undergone external validation but suggested that AI has potential associations between sperm characteristics and assisted reproductive outcomes.

AI in the analysis of testicular histopathology

Gadd et al. combined a pixel classifier with a convolutional neural network for the segmentation and classification of immunofluorescence sections, shortening the analysis time and retaining the biological relevance of tubular area and cell counts while artificial annotation is prone to losing this correlation (27). Yang et al. established a radiomics model based on multi-center magnetic resonance imaging (MRI), integrating multiple parameter sequences, for preoperative differentiation between seminoma and non-seminoma (28). These studies indicate that AI has application prospects in testicular histopathology, but it is still in the early clinical stage.

Generative AI helps reproductive scientists learn quickly

By quickly searching and summarizing a large amount of biomedical literature, LLMs can assist in evaluating the clinical significance of genetic variations in idiopathic MI. Moreover, when combined with knowledge graphs for cross-domain data mining, LLMs provide new hypotheses on the pathophysiological mechanisms of idiopathic infertility. However, given the risks of hallucinations and fabricated citations in large models, strict manual assessment and supervision must be maintained during clinical translation (29). One study developed a custom model called UroGPT, trained according to European urological guidelines. UroGPT was compared with ChatGPT4o by asking clinical questions, and two experts reviewed the consistency of the responses with the guidelines. The results showed that UroGPT’s responses were more in line with the guidelines, but it can only be used in academic research, not directly in clinical practice (30).

ED: multi-angle AI engagement from screening to treatment

ED is defined as the inability to attain or maintain a penile erection sufficient for satisfactory sexual performance (31). Traditional diagnosis relies on the International Index of Erectile Function Questionnaire, which is highly subjective and struggles to reveal underlying causes (32). AI intervention attempts to overcome this traditional dilemma by providing early warnings of risk, objective evaluations and decision support.

Risk stratification and clinical decision support

In terms of risk, early warning and stratification, AI has the potential to integrate multidimensional clinical data to make personalized predictions. Studies confirm that ED is a precursor to cardiovascular diseases. Belladelli et al. used the Chi-squared Automatic Interaction Detection algorithm to analyze the clinical data of patients with ED, identifying five clusters with varying risks of vasculogenic ED and three significant independent predictors: age, BMI and smoking status. The ‘high-risk’ group consisted of smokers aged under 53 years with a BMI over 25 kg/m², while the ‘very high-risk’ group comprised smokers aged over 53 years. The study noted that these two groups were more likely not to respond to phosphodiesterase type 5 inhibitors, and that the ‘very high-risk’ group faced a 30% increased risk of experiencing a major cardiovascular event within the next 10 years (33). Chen et al. developed a clinical decision support system based on electronic health records. This system combines genetic algorithms with support vector machine (SVM) to select key features from age, comorbidities and associated variables for modelling. This system could be integrated into hospital information systems in the future to support clinical risk assessment and preventive intervention. The study had a large sample size and built an online clinical decision-making system that could be deployed in a real environment in the future. However, laboratory data are lacking and limited to a single region (34).

Concurrently, ED is a common complication of radical prostatectomy (35). Jiang et al. retrospectively collected multidimensional data on 1,147 patients, including demographics, clinical and tumor characteristics, as well as operative and postoperative metrics, and trained and compared four mainstream ML algorithms: extreme gradient boosting (XGBoost), RF, SVM and k-Nearest Neighbors. The research team found that XGBoost performed the best in all evaluations. Therefore, this model was used as a reliable tool for personalized postoperative management (36).

Wearable devices and private home monitoring

The traditional gold standard for diagnosing ED is the nocturnal penile tumescence test. This test relies on clinical equipment such as RigiScan. However, these devices are large and can generate active mechanical loads, which can easily interfere with sleep and spontaneous erections. As a result, patient compliance decreases and diagnostic accuracy is affected. To address these issues, AI is driving the development of more portable diagnostic devices that are more suitable for home use (37). Wang et al. developed a wearable adaptive hardness monitoring system. Sng et al. proposed a home-based wearable quantitative monitoring device called EDSEN. Both schemes can continuously and quantitatively assess penile erection and hardness through lightweight sensors. Preliminary validation shows that the results of these devices are consistent with those of clinical devices. Based on this consistency, patient compliance has been improved, and the privacy of diagnosis has been enhanced (38,39).

Revealing the cause of the disease and targeted therapy

Micro-robot technology has demonstrated significant potential for transformation in the precise targeted treatment of ED. A research team has recently developed an intelligent magnetic soft micro-robot for the treatment of ED. In the ED models of mice and beagle dogs, the micro-robots successfully promoted the recovery of erectile function, providing a new interdisciplinary medical-engineering paradigm for targeted intervention therapy of ED (40).

The application of AI in medical image analysis provides a new approach for the etiological diagnosis of ED. Male erection is mainly regulated by the brain. Li et al. conducted a study using multimodal MRI and ML methods. They found that there were changes in cortical volume and white matter in patients with venous ED, which were associated with clinical symptoms such as anxiety, depression, and sexual dysfunction. Although there are some shortcomings, such as a small sample size, this study showcases the potential of AI as an objective tool for elucidating the neural mechanisms of ED and assisting in imaging diagnosis (41).

LLMs offer online consultation services.

LLMs are reshaping the way ED patients access medical information, but their clinical reliability remains limited. Pan et al. evaluated model responses to common ED questions and found that the generated texts were often difficult to read and lacked clear guidance on interventions (42). Şahin et al. compared five mainstream chatbots using validated scales and confirmed that although Bard performed well in terms of readability, none of the models fully met the level of comprehension of the average patient (43). This suggests that simplified language needs to be tailored to the patient before LLMs can be used in the clinical consultation of ED patients.

PE: integrating neuroimaging, metabolomics and generative AI

PE is a common male sexual dysfunction. Traditional etiological discussions have predominantly focused on psychological, behavioral and peripheral neural factors. AI is being applied in PE research in two ways: as an interactive information tool for patients and as a complex data analysis tool for researchers. Together, these applications reveal the potential and challenges of AI in understanding and managing PE.

Advances in neuroimaging techniques have led to a growing focus in research on central neural mechanisms in the brain. Lifelong PE (LPE) is defined as a condition in which ejaculation always or almost always occurs before or within approximately one minute of vaginal penetration from the first sexual experience (44). Xu et al. investigated the neural mechanisms of LPE based on resting-state functional connectivity. Using ML, they identified key functional connectivity that most effectively distinguished LPE patients. These connections primarily involved regions such as the medial orbitofrontal cortex, insula, globus pallidus and temporal lobe, providing specific targets for understanding PE-related brain network dyssynergia (45). Brain functional connectivity is not static, and its dynamic characteristics over time may better reflect the nature of the disease. Lu et al. introduced a study combining dynamic functional connectivity analysis with ML to successfully distinguish brain network states and transition patterns specific to LPE patients. The researchers found that patients spent prolonged periods in certain network states and exhibited abnormal transition frequencies between states. This offers a novel neurodynamic perspective for PE research (46).

The neural basis of LPE may involve multifaceted alterations to brain structure, white matter microstructure and functional activity. Geng et al. integrated structural MRI, diffusion tensor imaging and functional MRI data and analyzed them using a multiple kernel SVM. Their results indicated that LPE is not caused by abnormalities in a single brain region, but rather by functional dysregulation across multiple brain networks, including the default mode network, affective network, sensorimotor network, reward circuit and frontoparietal control network (47). However, these three studies are still in the exploratory phase and are limited by small sample sizes. Larger cohort studies are therefore needed in the future.

Omics research and prediction of precision medicine treatment

In addition to relying on macroscopic neuroimaging features, ML has also made significant breakthroughs in identifying the microscopic molecular biomarkers of PE and guiding precise medication. Luan et al. conducted a prospective multicenter study that combined plasma metabolomics with three ML algorithms: least absolute shrinkage and selection operator (LASSO) regression, SVM, and RF. This study successfully constructed a metabolic feature model for objectively diagnosing LPE and further used these metabolic profiles to precisely predict patients’ responses to dapoxetine treatment. This has opened a new path for individualized drug treatment for male sexual dysfunction (48).

An increasing number of patients are using LLMs to obtain health information due to their privacy and portability. However, recent standardized evaluations have shown consistency issues. In a comparative study of 25 common questions related to PE, Llama had the highest information quality score, and Llama and Gemini were superior to ChatGPT in terms of reliability, but all three generated texts that were beyond the comprehension of ordinary patients. Other studies also support the above findings and point out that ChatGPT content has problems with unreliable source citations and overly complex texts. LLMs can be used as a preliminary consultation tool, but the existing consensus believes that their information quality and readability have not yet reached the standard of reliable medical advice, and they cannot replace professional consultation and individualized treatment (49-51).

Prostatic diseases: discovery of biomarkers to support clinical decision making

The causes of prostatic diseases are quite complex, and their clinical manifestations also show diverse characteristics. Under such circumstances, the traditional diagnostic and treatment models often encounter bottlenecks. After the introduction of AI technology, the systematic management of prostatic diseases has gained a new conceptual framework. This framework can help clarify the pathogenesis of the disease, enhance the accuracy of the diagnostic process, and further optimize clinical decisions.

Exploring mechanisms and identifying biomarkers

The causes of chronic prostatitis/chronic pelvic pain syndrome (CP/CPPS) are rather complex, and the heterogeneity of symptoms is also obvious. When using traditional methods to diagnose and analyze the phenotype of this disease, considerable difficulties are encountered. During the exploration of the disease mechanism, AI technology can integrate multi-dimensional data and thereby reveal the pathological correlations that traditional methods are unable to capture (52). The pathogenesis of CP/CPPS is not limited to peripheral organs. The functional remodeling of the central nervous system is also included (53). One study compared the differences in brain functional connectivity between the patient group and the healthy control group and observed that the resting-state functional connectivity of the patients showed a weakened state. These abnormal connectivity features were subsequently used as input data to build an SVM classification model. The establishment of this model provided imaging evidence for the central nervous system pathological mechanism of CP/CPPS (54).

Beyond neuroimaging, constructing predictive models using easily accessible clinical and lifestyle data is essential for achieving disease management. One study innovatively combined dietary pattern analysis with ML, extracting two main dietary patterns—one rich in red meat and processed foods, and the other rich in dairy products — from food frequency questionnaire data. After feature selection via LASSO regression, they found that the red meat and processed food pattern was significantly associated with the severity of CP/CPPS symptoms. This provides data support for implementing individualized nutritional interventions (55). Another study utilized integrated bioinformatics and multiple ML algorithms to successfully identify and validate key candidate genes as biomarkers for the early diagnosis and prognosis assessment of prostate cancer (PCa) from multi-center gene expression data, thus providing potential targets for immunotherapy (56).

Radiomics study of non-invasive diagnosis and pathological classification

In the field of diagnosis and differential diagnosis, AI is driving the development of non-invasive, precise diagnoses by extracting detailed information from medical images. As distinguishing PCa from benign prostatic hyperplasia (BPH) remains a clinical challenge, Liu et al. used a radiomics-based ML predictive model to verify the immunohistochemical status of prostate tissue samples, exploring a new method for the non-invasive diagnosis of PCa (57). The research by Zhang et al. further revealed that incorporating a small amount of peritumoral tissue into PCa diagnoses can enhance the accuracy of deep learning models in distinguishing PCa from BPH (58). All the above studies are single-center retrospective analyses, and further external verification is needed.

AI can also assist with pathological subtyping to guide precision pharmacology. One study non-invasively predicted the pathological subtypes of BPH through MRI radiomics to guide the precise administration of 5α-reductase inhibitors, such as finasteride. Prospective validation showed that patients with gland-dominant hyperplasia experienced a significantly higher prostate volume reduction rate after taking finasteride than those with stroma-dominant hyperplasia, suggesting a new approach to BPH treatment (59).

The predictive model for surgical plans

In terms of predicting surgical difficulty and personalized planning, ML models can effectively evaluate surgical complexity and assist in selecting procedures by integrating clinical and imaging features. For example, one study constructed a model using light gradient boosting machine (LightGBM) algorithms to measure MRI features such as the prostate’s morphological angles in four directions, intravesical prostatic protrusion and mural nodules, combined with clinical metrics such as age and prostate volume. This is beneficial for the preoperative assessment of transurethral thulium laser enucleation of the prostate (ThuLEP) (60).

LLMs assist in clinical decision-making

LLMs are emerging tools that demonstrate potential in knowledge retrieval and case analysis. One study evaluated two LLMs for BPH clinical support and found both adequate in knowledge tests and mock diagnostics, but weak in complex cases (61). In PCa research, AI was tested with multiple-choice and real cases, with blinded scoring by senior doctors. Based on the results, DeepSeek-R1’s performance is closer to the requirements of clinical guidelines and is more in line with the logic of evidence-based reasoning. In contrast, ChatGPT o3 performs better in terms of language fluency and clarity of the workflow. However, both are limited to structured tasks and cannot yet replace doctors in practice (62).

Discussion

Differences in the clinical maturity of AI tools

The AI tools presented here cover a wide range of domains, and recognition of this gradient is critical to avoid premature adoption of unproven tools or to avoid excessive skepticism of mature applications (Table 2). The AI sperm fragmentation assessment system and the LensHooke^® R10 sperm detection kit have been commercialized in clinical laboratories, and the multi-dimensional sperm morphology analysis algorithm has been validated in multiple centers (11,14,17). The prediction models of ED after radical prostatectomy, ThuLEP difficulty, and micro-TESE have been externally or independently validated, and show robust performance, which are near to clinical translation (24,36,60). MRI-based radiomics models for distinguishing BPH from PCa and guiding finasteride treatment have achieved encouraging results in prospective studies, which need to be confirmed in multiple centers (59).

In the middle layer, neuroimage-based LPE classifiers, CP/CPPS for brain functional connectivity characterization, and Raman spectral-based sperm selection for assisted reproductive technology represent methodological innovations that have achieved high internal accuracy but are still in the proof-of-concept stage (25,45-47,54). Bioinformatics studies integrating multi-omics and ML, although computationally complex, have not advanced beyond the validation of computer simulations (19,56). Wearable devices for nocturnal erectile monitoring have demonstrated their feasibility in volunteers but require rigorous clinical testing in the target population (38,39). Notably, the metabolomics-based diagnostic and therapeutic prediction model for PE developed by Luan et al., although derived from a prospective multicenter design, still awaits independent external validation (48).

LLMs can be used for anonymous initial screening and information acquisition and play an auxiliary role in male diseases. However, they are not readable, the source is unreliable, and they cannot cope with high-risk problems, so it is more suitable as a research aid (Table 3). The micro-robots are still in the animal experiment stage, and the human experiment has not been started (40).

Pathways to clinical integration and adoption barriers

For the andrologist, the integration of AI in the clinical practice of andrology is gradually running through three stages: risk screening and patient education before treatment, auxiliary diagnosis and decision support during treatment, efficacy monitoring and follow-up management after treatment. Despite the rapid accumulation of AI research in andrology, the clinical integration of these tools remains constrained by several interrelated barriers. These can be grouped into three categories: regulatory approval, reimbursement and cost-effectiveness, and validation in routine practice.

In terms of regulatory clearance, many countries have established AI medical regulatory frameworks, but there are some problems in the implementation, such as a lack of evaluation standards and implementation rules, and insufficient connection between regulatory approval and medical insurance reimbursement. Moreover, regulatory agencies, ethics committees and clinicians lack systematic training, and the ethics and governance framework is difficult to truly implement (63-65). At the reimbursement level of medical insurance, the medical insurance payment of AI medical devices still relies on the pay-per-use framework, and there is a lack of exclusive categories suitable for workflows. The disconnect between approval and clinical value restricts sustainable promotion in routine care (66). The literature points out that software costs are classified as indirect expenses by the current payment system, and hospitals lack direct compensation for AI procurement. The use of AI may also lead to additional diagnostic findings, driving up health care expenditures. This leaves a payment system that both discourages AI adoption and potentially increases costs for use (67,68). At the level of practical validation, AI research in the field of male health is currently dominated by retrospective designs, with small sample sizes and mostly limited to single centers and internal validations. These studies are prone to recall bias and interference from various confounding factors, and long-term follow-up data are also scarce. There is a high risk of overfitting in the models, and the transparency of their decision-making process is insufficient. Most studies adopt observational designs, which can only reveal the correlations between variables but cannot confirm direct causal relationships (69).

Prospects of AI applications in andrology

The limitations are interconnected and progressive. Future efforts must involve conducting multicenter, large-sample, prospective studies; establishing standardized protocols for data acquisition and processing; developing more interpretable and robust algorithms; carrying out rigorous, multi-tiered external validations; and building mature ethical regulatory frameworks. Currently, advancements in computer hardware provide robust support for AI. Combining hardware with algorithms is essential to enhance model interpretability and robustness, thereby facilitating the application of AI in andrology (70). Concurrently, how andrologists can effectively utilize AI in diagnosis and treatment to achieve synergistic effects will also become a key area of research.

Conclusions

AI has been applied in multiple subfields of andrology. These subfields include MI, ED, PE, and prostatic diseases. In terms of detection, imaging, intelligent devices, disease prediction, and early screening, AI demonstrates certain advantages. However, the current evidence base is mainly retrospective and single-center studies, and external validation is limited. Only a few AI tools have progressed to multicenter or prospective clinical testing. Rigorous external validation, standardized data acquisition protocols, and improved model interpretability remain prerequisites for the responsible translation of AI into routine andrology. At the same time, it is necessary to improve the regulatory system and the medical insurance reimbursement system. Andrology clinicians should also correctly and reasonably use AI to improve the patient’s medical experience and optimize the treatment plan.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0474/rc

Peer Review File: Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0474/prf

Funding: This work was supported by Jiangsu Provincial Science and Technology Plan Special Project (No. BK20231379); Key Project of the Jiangsu Province Traditional Chinese Medicine Innovation Center for Clinical Medicine in Obstetrics, Gynecology, and Reproduction (No. k2021j18-4); and Jiangsu Province Graduate Student Research and Practice Innovation Program Project (No. SJCX24_0966).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0474/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Jungwirth D, Haluza D. Artificial Intelligence and Public Health: An Exploratory Study. Int J Environ Res Public Health 2023;20:4541. [Crossref] [PubMed]
2. Hamet P, Tremblay J. Artificial intelligence in medicine. Metabolism 2017;69S:S36-40. [Crossref] [PubMed]
3. Jiang S, Li H, Zhang L, et al. Generic Diagramming Platform (GDP): a comprehensive database of high-quality biomedical graphics. Nucleic Acids Res 2025;53:D1670-6. [Crossref] [PubMed]
4. Jordan MI, Mitchell TM. Machine learning: Trends, perspectives, and prospects. Science 2015;349:255-60. [Crossref] [PubMed]
5. He J, Baxter SL, Xu J, et al. The practical implementation of artificial intelligence technologies in medicine. Nat Med 2019;25:30-6. [Crossref] [PubMed]
6. Stiglic G, Kocbek P, Fijacko N, et al. Interpretability of machine learning-based prediction models in healthcare. WIREs Data Mining Knowl Discov 2020;10:e1379.
7. Garcea F, Serra A, Lamberti F, et al. Data augmentation for medical imaging: A systematic literature review. Comput Biol Med 2023;152:106391. [Crossref] [PubMed]
8. Wu X, Tao C, Li Z, et al. Medical image segmentation data augmentation method based on channel weight and data-efficient features. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi 2024;41:220-7. [Crossref] [PubMed]
9. Dao L, Liu N, Zheng G, et al. EALMDA:a Data Augmentation Method for MedicalNamed Entity Recognition. Journal of Zhengzhou University 2026;58:43-50. (Natural Science Edition).
10. Liu K, Mao X, Pan F, et al. Correlation analysis of sperm DNA fragmentation index with semen parameters and the effect of sperm DFI on outcomes of ART. Sci Rep 2023;13:2717. [Crossref] [PubMed]
11. Yang H, Ma M, Chen X, et al. Multidimensional morphological analysis of live sperm based on multiple-target tracking. Comput Struct Biotechnol J 2024;24:176-84. [Crossref] [PubMed]
12. Chen Y, Hu X, Liu Y. Advances in the Application of Artificial Intelligence in the Field of Male Infertility. Sichuan Da Xue Xue Bao Yi Xue Ban 2025;56:563-70. [Crossref] [PubMed]
13. Shahali S, Mortimer ST, McLachlan R, et al. Refined trajectory smoothing and deep learning classification of human sperm motility. Hum Reprod 2026;41:332-42. [Crossref] [PubMed]
14. Lin SB, Wang GQ, Li P. Artificial intelligence fluorescence method versus traditional flow cytometry for detection of sperm DFI in oligospermia patients. Zhonghua Nan Ke Xue 2025;31:115-20.
15. Noy L, Barnea I, Mirsky SK, et al. Sperm-cell DNA fragmentation prediction using label-free quantitative phase imaging and deep learning. Cytometry A 2023;103:470-8. [Crossref] [PubMed]
16. Hsu CT, Lee CI, Huang CC, et al. Development and integration of LensHooke(®) R10 for automatic and standardized diagnosis for sperm DNA fragmentation. Andrology 2023;11:1337-44. [Crossref] [PubMed]
17. Kuroda S, Karna KK, Kaiyal RS, et al. Novel sperm chromatin dispersion test with artificial intelligence-aided halo evaluation: A comparison study with existing modalities. Andrology 2023;11:1581-92. [Crossref] [PubMed]
18. Kobayashi H, Uetani M, Yamabe F, et al. A new model for determining risk of male infertility from serum hormone levels, without semen analysis. Sci Rep 2024;14:17079. [Crossref] [PubMed]
19. Gong ZZ, Feng QJ, Gao QH, et al. Shared mechanisms of obesity and male infertility and prediction of natural therapeutics based on machine learning and single-cell sequencing. Zhongguo Zhong Yao Za Zhi 2026;51:1766-74. [Crossref] [PubMed]
20. Parekattil SJ, Gudeloglu A. Robotic assisted andrological surgery. Asian J Androl 2013;15:67-74. [Crossref] [PubMed]
21. Douroumis K, Spartalis E, Stravodimos K, et al. Robotic-assisted microsurgery in andrology: a systematic review. Asian J Androl 2023;25:454-61. [Crossref] [PubMed]
22. Mohamed M, Yuriko E, Kawagoe Y, et al. AI-Based Augmented Reality Microscope for Real-Time Sperm Detection and Tracking in Micro-TESE. Bioengineering (Basel) 2026;13:102. [Crossref] [PubMed]
23. Kandil H, Arafa M, Kaya C, et al. Artificial Intelligence in Non-Obstructive Azoospermia: State of the Art in Sperm Retrieval. World J Mens Health 2026; Epub ahead of print. [Crossref]
24. Gül M, Şahin A, Özer C, et al. Machine Learning-Based Prediction of Sperm Retrieval Outcomes in Patients With Klinefelter Syndrome: A Multicenter Study With External Validation. Andrology 2026; Epub ahead of print. [Crossref]
25. Gao B, Zhang N, Bi D, et al. Label-free Raman spectroscopy combined with artificial intelligence for functional subtyping of human sperm and prediction of embryo development outcomes. Spectrochim Acta A Mol Biomol Spectrosc 2026;352:127498. [Crossref] [PubMed]
26. Li H, Gao J, Li Y. Machine learning-based prediction of IVF/ICSI outcomes in male factor infertility highlighting couple-level BMI. Front Endocrinol (Lausanne) 2026;17:1772106. [Crossref] [PubMed]
27. Gadd AJR, Sanou I, Brain E, et al. Machine learning and automation methods for the segmentation, classification and quantification of testicular tissue sections. Reprod Fertil 2026;7:RAF250210. [Crossref] [PubMed]
28. Yang S, Yang J, Xiao W, et al. Machine learning of multiparametric MRI radiomics preoperatively distinguishes testicular seminoma from non-seminoma: a multicenter study. Abdom Radiol (NY) 2026; Epub ahead of print. [Crossref]
29. M Jamil H. Future directions in infertility research: the role of generative AI and large language models. Syst Biol Reprod Med 2026;72:185-210. [Crossref] [PubMed]
30. Kaplan-Marans E, Katlowitz YE, West M, et al. From ChatGPT to UroGPT: A guideline-trained artificial intelligence model for male infertility. Curr Urol 2026;20:135-40. [Crossref] [PubMed]
31. Teoman AS, Serefoglu EC. Artificial Intelligence-Based Clinical Decision-Making in Erectile Dysfunction: a Narrative Review. Curr Urol Rep 2024;26:22. [Crossref] [PubMed]
32. Du GW, Niu PN, Yang ZX, et al. Association between metabolic parameters and erection in erectile dysfunction patients with hyperuricemia. Asian J Androl 2025;27:482-7. [Crossref] [PubMed]
33. Belladelli F, Cei F, Pozzi E, et al. A novel algorithm-based risk classification for vascular damage in men with erectile dysfunction. J Sex Med 2025;22:291-7. [Crossref] [PubMed]
34. Chen YF, Lin CS, Hong CF, et al. Design of a Clinical Decision Support System for Predicting Erectile Dysfunction in Men Using NHIRD Dataset. IEEE J Biomed Health Inform 2019;23:2127-37. [Crossref] [PubMed]
35. Shen C, Jain K, Shah T, et al. Relationships between erectile dysfunction, prostate cancer treatment type and inflatable penile prosthesis implantation. Investig Clin Urol 2022;63:316-24. [Crossref] [PubMed]
36. Jiang H, Ji L, Zhu L, et al. XGBoost model for predicting erectile dysfunction risk after radical prostatectomy: development and validation using machine learning. Discov Oncol 2025;16:810. [Crossref] [PubMed]
37. Trip EJ, Elzevier HW, Pelger RCM, et al. Patients' perceptions of nocturnal erectile function assessment with the RigiScan®. Int J Impot Res 2025;37:832-6. [Crossref] [PubMed]
38. Wang X, Wang R, Zhang Y, et al. A wearable adaptive penile rigidity monitoring system for assessment of erectile dysfunction. Microsyst Nanoeng 2024;10:131. [Crossref] [PubMed]
39. Sng CMN, Wee LMC, Tang KC, et al. Wearable Soft Microtube Sensors for Quantitative Home-Based Erectile Dysfunction Monitoring. Sensors (Basel) 2022;22:9344. [Crossref] [PubMed]
40. Wang S, Wang Z, Shen Z, et al. Magnetic soft microrobots for erectile dysfunction therapy. Proc Natl Acad Sci U S A 2024;121:e2407809121. [Crossref] [PubMed]
41. Li L, Fan W, Li J, et al. Abnormal brain structure as a potential biomarker for venous erectile dysfunction: evidence from multimodal MRI and machine learning. Eur Radiol 2018;28:3789-800. [Crossref] [PubMed]
42. Pan A, Musheyev D, Loeb S, et al. Quality of erectile dysfunction information from ChatGPT and other artificial intelligence chatbots. BJU Int 2024;133:152-4. [Crossref] [PubMed]
43. Şahin MF, Ateş H, Keleş A, et al. Responses of Five Different Artificial Intelligence Chatbots to the Top Searched Queries About Erectile Dysfunction: A Comparative Analysis. J Med Syst 2024;48:38. [Crossref] [PubMed]
44. Waldinger MD. The pathophysiology of lifelong premature ejaculation. Transl Androl Urol 2016;5:424-33. [Crossref] [PubMed]
45. Xu Z, Yang X, Gao M, et al. Abnormal Resting-State Functional Connectivity in the Whole Brain in Lifelong Premature Ejaculation Patients Based on Machine Learning Approach. Front Neurosci 2019;13:448. [Crossref] [PubMed]
46. Lu J, Chen Q, Li D, et al. Reconfiguration of Dynamic Functional Connectivity States in Patients With Lifelong Premature Ejaculation. Front Neurosci 2021;15:721236. [Crossref] [PubMed]
47. Geng B, Gao M, Piao R, et al. Multivariate Pattern Analysis of Lifelong Premature Ejaculation Based on Multiple Kernel Support Vector Machine. Front Psychiatry 2022;13:906404. [Crossref] [PubMed]
48. Luan J, Shao W, Xu C, et al. A prospective multicenter study on plasma metabolomics and machine learning for diagnosing premature ejaculation and predicting dapoxetine treatment response. J Sex Med 2026;23:qdag026. [Crossref] [PubMed]
49. Anıl H, Kayra MV. The digital dialogue on premature ejaculation: evaluating the efficacy of artificial intelligence-driven responses. Int Urol Nephrol 2025;57:2829-36. [Crossref] [PubMed]
50. Gonultas S, Kardas S, Gelmis M, et al. Chatbots' performance in premature ejaculation questions: a comparative analysis of reliability, readability, and understandability. Int J Impot Res 2025;37:843-9. [Crossref] [PubMed]
51. Şahin MF, Keleş A, Özcan R, et al. Evaluation of information accuracy and clarity: ChatGPT responses to the most frequently asked questions about premature ejaculation. Sex Med 2024;12:qfae036. [Crossref] [PubMed]
52. Ma X, Lao Y, Bai Y, et al. Study progress of etiologic mechanisms of chronic prostatitis/chronic pelvic pain syndrome. Int Immunopharmacol 2025;148:114128. [Crossref] [PubMed]
53. Clemens JQ, Mullins C, Ackerman AL, et al. Urologic chronic pelvic pain syndrome: insights from the MAPP Research Network. Nat Rev Urol 2019;16:187-200. [Crossref] [PubMed]
54. Lan X, Niu X, Bai WX, et al. The functional connectivity of the basal ganglia subregions changed in mid-aged and young males with chronic prostatitis/chronic pelvic pain syndrome. Front Hum Neurosci 2022;16:1013425. [Crossref] [PubMed]
55. Wang Z, Wu W, Wen B, et al. Dietary patterns and chronic prostatitis: a symptom severity prediction model based on nutritional clustering and machine learning. Front Nutr 2025;12:1660430. [Crossref] [PubMed]
56. Yuan J, Zhou D, Yu S. Interpretable machine learning driven biomarker identification and validation for prostate cancer. Transl Androl Urol 2025;14:1528-41. [Crossref] [PubMed]
57. Liu YF, Shu X, Qiao XF, et al. Radiomics-Based Machine Learning Models for Predicting P504s/P63 Immunohistochemical Expression: A Noninvasive Diagnostic Tool for Prostate Cancer. Front Oncol 2022;12:911426. [Crossref] [PubMed]
58. Zhang Y, Li W, Zhang Z, et al. Differential diagnosis of prostate cancer and benign prostatic hyperplasia based on DCE-MRI using bi-directional CLSTM deep learning and radiomics. Med Biol Eng Comput 2023;61:757-71. [Crossref] [PubMed]
59. Wang J, Tang W, Zhu J, et al. Predicting the pathological subdiagnosis of benign prostatic hyperplasia with MRI radiomics: A noninvasive approach. View 2025;6:20240092.
60. Yang S, He S, Li J, et al. MRI-based morphologic parameter model for preoperative prediction of transurethral thulium laser enucleation of the prostate difficulty. Quant Imaging Med Surg 2026;16:237. [Crossref] [PubMed]
61. Li Y, Meng C, Peng L, et al. Real-world feasibility of generative large language models for clinical decision support in benign prostatic hyperplasia. Int J Surg 2026;112:2546-53. [Crossref] [PubMed]
62. Li R, Zhao A, Peng L, et al. Evaluating the Clinical Competence of Large Language Models in Prostate Cancer Management: A Comparative Study of DeepSeek-R1 and ChatGPT. Ann Surg Oncol 2026;33:1858-69. [Crossref] [PubMed]
63. Karako K, Tang W. Clinical artificial intelligence (AI) in Japan: Regulatory pathways, domain-specific evidence, and its data infrastructure from an international perspective. Glob Health Med 2026;8:166-81. [Crossref] [PubMed]
64. Chung J, Shin Y, Kim G, et al. The Enactment and Implementation of Korea's Digital Medical Products Act: A Comparative Study Exploring Policy Implications. J Korean Med Sci 2026;41:e155. [Crossref] [PubMed]
65. Mwaka J, Mulebeke R, Dedibo R, et al. Capacity building for ethical use of artificial intelligence in health: protocol for a scoping review of training initiatives and gaps in Africa. BMJ Open 2026;16:e111660. [Crossref] [PubMed]
66. Bains SS, Mohan V, Gold JA. Translating innovation into payment: A.I. reimbursement in U.S. health care. J Am Med Inform Assoc 2026;33:1410-4.
67. Neprash HT. Medicare Spending On Artificial Intelligence: Payment Policy Is Only Part Of The Story. Health Aff (Millwood) 2026;45:26-8. [Crossref] [PubMed]
68. Longyear RL, Berenson RA. Artificial Intelligence Payment Policies: Challenges For CMS And The Medicare Physician Fee Schedule. Health Aff (Millwood) 2026;45:14-21. [Crossref] [PubMed]
69. Clark H, Venishetty N, Howell S, et al. Utilization of artificial intelligence in Men's Health: Opportunities for innovation and quality improvement. Int J Impot Res 2026;38:381-92. [Crossref] [PubMed]
70. Berggren K, Xia Q, Likharev KK, et al. Roadmap on emerging hardware and technology for machine learning. Nanotechnology 2021;32:012002. [Crossref] [PubMed]